Skip to main content

October 8, 2026

Highlights​

A lot landed this release. Here is the short version:

  • Everyone: tell us (and your agents) what is working. You can now give a thumbs up or down on AI replies in chat, and agent owners get a new Feedback tab in the agent editor to see how their agents are doing over time. Chat also gets smoother, with a new way to mention agents and clearer guidance when you want to create or edit one.
  • Builders: GitHub is now part of the conversation. Connect GitHub from chat, pick one of your repositories, and let the AI work on the code and open a pull request for you. Importing a GitHub project into App Builder now brings in the real project history too.
  • Org admins: fewer interruptions from empty balances. New auto-reload settings can top up chat credits for you when they run low, and the new Claude Haiku 5.5 model is available in the model catalog.

There is more in the full release notes, including sign-in reliability fixes and polish across chat and App Builder.

New features​

Rate replies and see how your agents are doing​

Good feedback makes good agents. You can now give a thumbs up or down on assistant replies right from the message actions in chat, so you can quickly flag what hit the mark and what missed.

If you own an agent, the new Feedback tab in the agent editor shows how ratings change over time, so you can spot improvements or regressions and fine-tune your agent with real signals instead of guesswork. Ratings follow your agent if it moves when organizations are merged.

Work with GitHub right from chat​

You no longer need to leave the conversation to work with your code. When GitHub is not connected yet, chat offers a connect card, and when the app still needs to be installed, an install card walks you through it.

Once you are connected, the AI can:

  • List the repositories you have access to
  • Pull a repository into a chat so it can read and change the code
  • Open a pull request with the changes, even when you are chatting without a dedicated agent

Your GitHub access stays scoped to your own account and is only used for the work you ask for.

In App Builder, importing a project from GitHub now brings in the real repository history, so your imported app keeps its context.

Mention agents right in your message​

Type to mention an agent and pick it from a suggestion list. Mentions show up as clean pills in your message, so it is clear who you are talking to. This is rolling out gradually, so not every workspace will see it yet.

Create and edit agents from chat, more clearly​

Chat now makes it easier to edit an existing agent and no longer nudges you toward creating a new one when you did not ask for it. The AI makes focused changes to your agent's configuration, so edits are more predictable.

Auto-reload for chat credits​

Running out of credits in the middle of work is frustrating. Auto-reload lets organizations set up automatic top-ups when the balance runs low, so teams can keep working without waiting for an admin. If you turn it off and back on later, it starts a fresh reload cycle.

New model: Claude Haiku 5.5​

Claude Haiku 5.5 is now in the model catalog, giving you another fast option for everyday chats and agents.

Improvements​

  • Sign-in is steadier, including a fix for Safari redirect loops and better recovery when your marketplace session needs to be refreshed
  • More reliable chat and sandboxes, with better handling of long outputs, uploads, and files that were left behind
  • Clearer connector warnings that show once and go away when the connector becomes usable
  • Routines show agent names in their instructions so they are easier to read
  • Longer skill generation can finish without timing out
  • Stronger protection for your GitHub credentials while code runs in chat

Bug fixes​

General reliability improvements across chat, sign-in, App Builder imports, agent editing, and billing settings. Together they should make everyday work feel steadier.

⌘J